acceptodds
Under review as a conference paper at ICLR 2027

CFGP: Closed-Form Graph Prompting without Per-Task Training

Abstract

Few-shot graph learning typically proceeds by pre-training a graph neural network (GNN) on unlabeled graph data and adapting it to a labeled downstream task. Graph prompt learning freezes this encoder and adapts a small set of prompts instead of fine-tuning all parameters. Existing methods, however, still typically train prompt parameters for each task, which is costly and sensitive to initialization, learning rate, and early stopping under extreme label scarcity. We propose CFGP (Closed-Form Graph Prompting without Per-Task Training), which builds an instance-specific input prompt in closed form from class signatures derived from the labeled examples and a clustering of all instances, then reads it out with the frozen GNN and a prototype classifier. Adaptation uses no optimizer and no training objective. The same construction covers node- and graph-level tasks, differing only in how the prompt is injected, and cuts adaptation time by more than an order of magnitude at competitive accuracy.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
graph prompt learning, few-shot learning, closed-form prompting, training-free adaptation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_yeHnOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

CFGP replaces per-task prompt optimization with a prompt computed from support-set class prototypes and clustering assignments. The prompt modifies the input features, which are then re-encoded by a frozen GNN. The method supports both node and graph classification. Its main appeal is the reduced adaptation cost, with competitive results on several of the eight evaluated datasets.

Critical strengths

Removing the backward pass makes adaptation substantially cheaper. The construction is straightforward and works for both task types without training a separate prompt module.

Critical weaknesses
  1. My main concern is the use of evaluation tasks for configuration selection. Appendix A specifies a search over 162 configurations per dataset, evaluated on the same tasks used to report the final results. This is test-set tuning, even though the prediction procedure itself does not read test labels. The shared-configuration result is encouraging, but the paper needs to explain how that configuration was chosen and report results on tasks excluded from selection. The timing comparison should also account separately for configuration search.

  2. The baseline changes in Appendix E make several comparisons difficult to interpret. GPF, GPF+, All-in-One, and SUPT are applied to cached encoder outputs, changing their intended interaction with the GNN. These results do not establish an advantage over faithful implementations of those methods. I would also expect comparisons with simple training-free alternatives, particularly label propagation and transductive prototype refinement.

Questions for authors

See weaknesses

Reviewer_awgaOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes CFGP, a closed-form graph prompting method for few-shot adaptation with a frozen GNN. Instead of optimizing task-specific prompt parameters, CFGP combines class responsibilities from the support set with spherical k-means assignments over unlabeled instances to construct instance-specific input prompts, followed by frozen-GNN re-encoding. Experiments on eight node- and graph-classification datasets show substantially lower adaptation time than learned-prompt baselines, with competitive or better performance in many settings.

Critical strengths
  1. The paper addresses a meaningful inefficiency in few-shot graph prompting: existing methods still perform substantial per-task optimization despite having very limited supervision. CFGP provides a simple and structurally different alternative with much lower adaptation cost.

  2. The method is clearly specified and supported by useful ablations. The experiments suggest that clustering and instance-aligned prompt construction both contribute to the gains, and the balanced-accuracy analysis on graph datasets is informative.

Critical weaknesses
  1. The current experiments do not fully isolate the contribution of input-space prompting from the underlying transductive refinement procedure. CFGP combines support-derived class responsibilities with clustering over unlabeled query instances, and these components may themselves account for part of the gain. A controlled baseline that uses the same responsibilities and clustering information but performs refinement directly in representation space would help clarify how much of the improvement specifically comes from prompt injection followed by frozen-GNN re-encoding.

  2. The main CFGP results rely on dataset-specific hyperparameter selection using accuracy on the downstream evaluation tasks, since the benchmark provides no validation split. This makes the practical adaptation setting somewhat unclear: although CFGP avoids gradient-based prompt optimization, its reported performance still depends on configuration selection informed by evaluation performance. It would therefore be helpful to clarify how the method would be configured on a genuinely unseen dataset without access to evaluation labels.

Questions for authors
  1. Please see Weakness 1. Could the authors provide a representation-space refinement baseline using the same class responsibilities and clustering information, but without input prompt injection and GNN re-encoding?
  2. Please see Weakness 2. In particular, how would the hyperparameters be selected for a genuinely unseen dataset without access to downstream evaluation labels? How was the shared configuration reported in the paper chosen?
Reviewer_LtZtOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CFGP, a closed-form graph prompting method for few-shot graph learning that avoids per-task prompt optimization by constructing instance-specific prompts from class signatures and clustering over frozen GNN embeddings. The method is evaluated on both node-level and graph-level classification tasks under 1-shot and k-shot settings, with comparisons against several graph prompting baselines and additional ablation and efficiency studies.

Critical strengths
  1. The motivation is clearly presented, especially the concern that per-task prompt optimization can be unstable and inefficient under extreme label scarcity.
  2. The experimental evaluation is relatively comprehensive, covering both node- and graph-level classification, k-shot settings, efficiency comparisons, and ablation studies.
Critical weaknesses
  1. The method appears to rely on a strong but insufficiently validated assumption that frozen embeddings already exhibit meaningful same-class clustering, which may be particularly fragile in the 1-shot setting.
  2. The proposed approach seems closer to label-guided clustering-based feature construction than conventional prompt learning, since it does not learn task-adaptive prompt parameters.
  3. The main hyperparameters are selected using the evaluation tasks, and the method is sensitive to these choices, which weakens the credibility of the reported “7 out of 8 datasets” advantage.
  4. Raw accuracy may not accurately reflect the true performance of the methods on class-imbalanced graph classification datasets, making the corresponding evaluation less reliable.
  5. The efficiency comparison is not conducted under matched computational budgets, and having no trainable prompt parameters does not mean the method has negligible computational cost.
Questions for authors

See Weaknesses.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.